Causal Inference for Crowd Management

From Observation to Intervention

Amir Rafe�Civil & Environmental Engineering�Utah State University

amir.rafe@usu.edu    https://pozapas.github.io https://github.com/pozapas

Where We’re Going Today

• Core Concepts in Causal Inference
• Why Causality Matters in Crowds
• Applications in Crowd Management
• Pedestrian Safety
• Evacuations
• Collective Dynamics
• From Models to Interventions
• Discussion and Takeaways

Why Causality? Why Now?

Causality is not just about data, it’s about

understanding and action.

​

“Data are profoundly dumb about causes and

effects.”— Judea Pearl, The Book of Why

​

We ask causal questions every day:

• “Will this intervention help?”
• “Would things have gone differently?”
• “What caused this to happen?”

Image Source https://www.penguin.com.au/books/the-book-of-why-9780141982410

The Ladder of Causation

Three Rungs of Causal Reasoning:

• Seeing (Association)

What is? What goes with what?

Example: People run more when alarms sound.

• Doing (Intervention)

What if we make X happen?

Example: Will adding an exit reduce evacuation time?

• Imagining (Counterfactuals)

What if we had done something differently?

• Example: Would that person have exited faster with staff guidance?

Image Source: Pearl, Judea, and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018

Correlation ≠ Causation

🔗 Correlation = Two variables move together

🧭 Causation = One variable changes because of the other

​

Common Causal Pitfalls:

• Confounding
• E.g., Ice cream sales ↑ and drowning ↑ (confounder: hot weather).
• Reverse causality
• E.g., More staff presence → safer crowd? Or unsafe crowd → more staff?
• Spurious correlation
• E.g., Elevator button pressed ≠ elevator comes faster

How We Think About Cause and Effect

Two Big Ideas in Causal Thinking:

1. What if things had been different?

→ Imagine someone exits in 30 seconds... would it have been faster if they saw

a staff member?�→ This is a counterfactual question.

​

2. What happens if we intervene?

→ If we add a new sign here, will people exit faster?�→ This is an intervention question.

​

​

Causal Diagrams (DAGs)

Do Thinking

Counterfactual Reasoning

Why Causality Matters in Crowds

• Crowds are complex systems
• People react to others, to the environment, and uncertainty
• Small changes (signs, doors, messages) can ripple widely

​

• Correlation-based decisions can mislead
• More exits ≠ faster evacuation
• More signs ≠ less crush or surging behavior
• More staff ≠ smoother flow

​

Why Causality Matters in Crowds

• Causality helps us ask:
• What if we change something?
• What would have happened otherwise?
• What intervention will actually work, and for whom?
• Real-world stakes:
• Lives lost or saved
• Resources wasted or optimized
• Equity achieved or overlooked

What Goes Wrong Without Causal Thinking?

• Common risks:
• Mistaking correlation for mechanism
• Acting on symptoms, not root causes
• Failing to anticipate unintended effects
• Misallocating resources, or even causing harm

Understanding Pedestrian Safety with Causal Tools

• Engineers rely on pedestrian data, but correlation-based models can mislead:
• People who cross mid-block might appear riskier...

→ but maybe they’re more experienced or time-pressured

• More traffic often → longer wait times

→ but do pedestrians actually wait because of traffic?

Understanding Pedestrian Safety with Causal Tools

Causal Insights from Research:

1. Debiased Machine Learning (DML)�📄 Kamal & Farooq, 2023 (https://doi.org/10.1177/03611981231152246)

• Estimated the causal effect of traffic density on pedestrian wait time
• Finding: Lower traffic density causally reduces wait time (−1.115 sec per unit drop)
• More precise than traditional copula models

✅ Policy Implication: Traffic calming = faster, more comfortable pedestrian flow

Causal Insights from Research:

1. Propensity Score Matching (PSM)�📄 Zhou et al., 2025 (https://doi.org/10.1016/j.aap.2025.107988)

• Assessed causal effect of weather (rain/fog) on perceived risk and pedestrian safety
• VR CAVE experiment and Controlled for other variables using matching
• Result: Perceived risk ↑ in rain/fog, especially at dusk

✅ Policy Implication: Add lighting + weather-sensitive warnings at crossings

Understanding Pedestrian Safety with Causal Tools

Evacuation: Moving from Simulation to Causal Insight

• Evacuation success isn’t just about exits and alarms
• It’s about what drives behavior
• Not all interventions work equally
• Wrong assumptions can delay response,

increase risk, or widen inequities

Evacuation: Moving from Simulation to Causal Insight

• What Triggers Action in Emergencies?

Common Assumption: "If people hear the alarm, they’ll evacuate faster.”

❌ But real-world behavior shows otherwise:

Alarm-only scenarios (e.g., beeping, sirens):

• People pause to film the scene
• Wait for confirmation from others
• Exhibit delayed movement or passive observation

“Is this real?” → “Let’s see what others do.”

Evacuation: Moving from Simulation to Causal Insight

How Building Design Shapes Evacuation Behavior

It’s not just exits and walls: design influences decisions.

• Fruin’s 4 Determinants of Pedestrian Movement:
• Space – Width, bottlenecks, clutter
• Time – Travel time, delay perception
• Information – Visibility, signage, clarity
• Energy – Effort required, slope, stairs

Evacuation: Moving from Simulation to Causal Insight

How Building Design Shapes Evacuation Behavior

​

• Exit Placement & Perceived Distance

→ People often choose familiar or visible exits, not the nearest

• Corridor Width

→ Affects comfort, perceived crowding, and movement speed

• Smoke & Fire Visibility

→ Triggers emotional responses: crush behavior, rerouting, freezing

• Psychological State

→ Fear, training, and familiarity shape decision pathways

Causality in Collective Behavior

Crowds are not collections of individuals; they are interactive systems.

• One person’s movement influences another’s
• Emotions, decisions, and trust spread through networks
• This violates standard assumptions like SUTVA (no interference between units)

Effective interventions must consider indirect effects

Causality in Collective Behavior

Direct effect:

You install a new exit in a stadium…

Indirect effects:

• People might reroute away from better exits due to visibility.
• The new exit might create bottlenecks elsewhere.
• It may shift crowd density and cause new risks near that zone.
• People with disabilities might still be unable to use it, worsening inequality.

Causal Features of Collective Behavior

• Social Contagion

Crush behavior spreads even without direct exposure to danger

Fearful movement → causes others to follow reflexively

“If they’re running, maybe I should too.”

​

• Leader–Follower Dynamics

Some agents guide others implicitly

Leadership may last only a few seconds

Directional influence can be quantified using: Transfer Entropy, Granger Causality

These tools estimate who causes whom to move, and when.

​

From Explanation to Action: Causal Models as Decision Tools

Why causal models matter beyond academia:

We don’t just want to know what happened, we need to predict:

“What would happen if we did X instead of Y?”

• Descriptive models = explain the past
• Predictive models = forecast based on patterns
• Causal models = simulate the effects of interventions

From Explanation to Action: Causal Models as Decision Tools

Causal Tools for Decision-Making

• do-calculus & DAGs

→ Test policy changes virtually

• Counterfactuals

→ Compare real vs. hypothetical outcomes

• Causal ABMs

→ Inject interventions into agent logic (e.g., AI-guided signage, smoke flow, training)

From Explanation to Action: Causal Models as Decision Tools

DAGs + do-calculus → Virtual Policy Testing

Should we add signage at a hallway junction to speed up evacuation?

From Explanation to Action: Causal Models as Decision Tools

Counterfactual Thinking → Ethical Evaluation

Would a disabled evacuee have exited faster if staff had intervened?

Real data: person took 90 seconds to evacuate unaided

Matched group with staff aid: median time = 45 seconds

Counterfactual question:

What would’ve happened if this person had received staff guidance?

→ Estimated counterfactual: Exit time = 52 seconds

✅ Justifies targeted deployment of trained staff for vulnerable individuals

Causal Thinking Is Essential for Crowd Management

• Correlation misleads.

→ Observing isn’t enough; we must ask what if?

• Crowd behavior is interdependent.

→ People influence each other; interventions have ripple effects.

• Causal models help us design, not just explain.

→ Tools like DML, DAGs, and counterfactuals can guide proactive interventions.

Let’s Reflect

• Where in your current research or practice might you be

relying on correlation?

• How might causal tools improve your evaluation of interventions?
• What challenges do you foresee in applying causal inference in your context?
• How can we better integrate disability and equity into causal models?

Social Media Engagement

Social Media Engagement

Thank You !

Human Behavior | ML & Applied AI | Causality in Engineering | Evacuation